gspo

Reproducing and studying RL algorithms for LLM agents, including GRPO, GSPO, DAPO, OPD, Search-R1, ReTool, ALFWorld and beyond.

Pure-Rust LLM runtime: one binary serves (OpenAI-compatible), runs local agents, and distills models on their own rollouts — on Apple Silicon and NVIDIA. No Python on the hot path.

Reinforcement learning for text generation on MLX (Apple Silicon)

Training and dataset-building pipeline for NVIDIA Nemotron reasoning experiments using Unsloth, TRL and NeMo RL.

Sync vs fully-async agentic RL on verl: multi-turn GRPO, long-tail rollout profiling, staleness ablations — quantifying when async pays off.

OPSG-based test refinement for Java: Stable RL approach to generate maintainable, high-quality unit tests with 98.6% compilation success.

Auditable online post-training lab for MiniMind: GRPO, CISPO, Dr.GRPO and GSPO with reproducible Math/Tool RLVR experiments.

RL agents across LangChain & LangGraph, 50+ MCP registry tools, PGVector stores, deployment w Dynamo and inference with vLLM & SGLang